delete old vectors as events removed from DB, add image similarity search

.devcontainer/devcontainer.json

@@ -10,7 +10,7 @@
   "features": {
     "ghcr.io/devcontainers/features/common-utils:1": {}
   },
-  "forwardPorts": [5000, 5001, 5173, 1935, 8000, 8554, 8555],
+  "forwardPorts": [5000, 5001, 5173, 1935, 8554, 8555],
   "portsAttributes": {
     "5000": {
       "label": "NGINX",
@@ -28,10 +28,6 @@
       "label": "RTMP",
       "onAutoForward": "silent"
     },
-    "8000": {
-      "label": "Chroma",
-      "onAutoForward": "silent"
-    },
     "8554": {
       "label": "gortc RTSP",
       "onAutoForward": "silent"

docker/main/Dockerfile

@@ -178,6 +178,10 @@ ENV NVIDIA_DRIVER_CAPABILITIES="compute,video,utility"
 
 # Turn off Chroma Telemetry: https://docs.trychroma.com/telemetry#opting-out
 ENV ANONYMIZED_TELEMETRY=False
+# Allow resetting the chroma database
+ENV ALLOW_RESET=True
+# Disable tokenizer parallelism warning
+ENV TOKENIZERS_PARALLELISM=true
 
 ENV PATH="/usr/lib/btbn-ffmpeg/bin:/usr/local/go2rtc/bin:/usr/local/nginx/sbin:${PATH}"
 
@@ -243,10 +247,10 @@ CMD ["sleep", "infinity"]
 FROM --platform=$BUILDPLATFORM node:20 AS web-build
 
 WORKDIR /work
-COPY web/package.json web/package-lock.json ./
+COPY web-old/package.json web-old/package-lock.json ./
 RUN npm install
 
-COPY web/ ./
+COPY web-old/ ./
 RUN npm run build \
     && mv dist/BASE_PATH/monacoeditorwork/* dist/assets/ \
     && rm -rf dist/BASE_PATH

frigate/app.py

@@ -31,6 +31,7 @@ from frigate.const import (
     MODEL_CACHE_DIR,
     RECORD_DIR,
 )
+from frigate.embeddings import Embeddings
 from frigate.embeddings.processor import EmbeddingProcessor
 from frigate.events.audio import listen_to_audio
 from frigate.events.cleanup import EventCleanup
@@ -341,6 +342,13 @@ class FrigateApp:
 
         migrate_db.close()
 
+    def init_embeddings(self) -> None:
+        self.embeddings: Embeddings = None
+        try:
+            self.embeddings = Embeddings(self.config)
+        except ValueError:
+            pass
+
     def init_go2rtc(self) -> None:
         for proc in psutil.process_iter(["pid", "name"]):
             if proc.info["name"] == "go2rtc":
@@ -407,6 +415,7 @@ class FrigateApp:
             self.onvif_controller,
             self.external_event_processor,
             self.plus_api,
+            self.embeddings,
         )
 
     def init_onvif(self) -> None:
@@ -590,7 +599,7 @@ class FrigateApp:
 
     def start_embeddings_processor(self) -> None:
         self.embeddings_processor = EmbeddingProcessor(
-            self.config, self.embeddings_queue, self.stop_event
+            self.config, self.embeddings, self.embeddings_queue, self.stop_event
         )
         self.embeddings_processor.start()
 
@@ -607,7 +616,7 @@ class FrigateApp:
         self.event_processor.start()
 
     def start_event_cleanup(self) -> None:
-        self.event_cleanup = EventCleanup(self.config, self.stop_event)
+        self.event_cleanup = EventCleanup(self.config, self.embeddings, self.stop_event)
         self.event_cleanup.start()
 
     def start_record_cleanup(self) -> None:
@@ -688,10 +697,11 @@ class FrigateApp:
             self.set_log_levels()
             self.init_queues()
             self.init_database()
+            self.bind_database()
+            self.init_embeddings()
             self.init_onvif()
             self.init_recording_manager()
             self.init_go2rtc()
-            self.bind_database()
             self.init_inter_process_communicator()
             self.init_dispatcher()
         except Exception as e:

frigate/embeddings/__init__.py

@@ -0,0 +1,103 @@
+"""ChromaDB embeddings database."""
+import base64
+import io
+import logging
+import os
+import time
+
+import numpy as np
+from chromadb import Collection
+from chromadb import HttpClient as ChromaClient
+from chromadb.config import Settings
+from PIL import Image
+from playhouse.shortcuts import model_to_dict
+
+from frigate.config import FrigateConfig
+from frigate.const import CONFIG_DIR
+from frigate.models import Event
+
+from .functions.clip import ClipEmbedding
+from .functions.minilm_l6_v2 import MiniLMEmbedding
+
+logger = logging.getLogger(__name__)
+
+
+def get_metadata(event: Event) -> dict:
+    """Extract valid event metadata."""
+    return {
+        k: v
+        for k, v in model_to_dict(event).items()
+        if k not in ["id", "thumbnail"]
+        and v is not None
+        and isinstance(v, (str, int, float, bool))
+    }
+
+
+class Embeddings:
+    """ChromaDB embeddings database."""
+
+    def __init__(self, config: FrigateConfig) -> None:
+        if not config.semantic_search.enabled:
+            raise ValueError("Semantic search is not enabled.")
+        self.config: FrigateConfig = config
+        self.client: ChromaClient = ChromaClient(
+            host="127.0.0.1",
+            settings=Settings(anonymized_telemetry=False),
+        )
+
+        reindex = os.path.exists(f"{CONFIG_DIR}/.reindex_events")
+        if reindex:
+            logger.info("Indexing event embeddings...")
+            self.client.reset()
+
+        self.thumbnail: Collection = self.client.get_or_create_collection(
+            name="event_thumbnail", embedding_function=ClipEmbedding()
+        )
+        self.description: Collection = self.client.get_or_create_collection(
+            name="event_description", embedding_function=MiniLMEmbedding()
+        )
+
+        if reindex:
+            # On startup, embed all existing events
+            st = time.time()
+
+            thumbnails = {"ids": [], "images": [], "metadatas": []}
+            descriptions = {"ids": [], "documents": [], "metadatas": []}
+
+            events = Event.select().where(
+                (Event.has_clip == True | Event.has_snapshot == True)
+                & Event.thumbnail.is_null(False)
+            )
+
+            event: Event
+            for event in events.iterator():
+                metadata = get_metadata(event)
+                thumbnail = base64.b64decode(event.thumbnail)
+                img = np.array(Image.open(io.BytesIO(thumbnail)).convert("RGB"))
+                thumbnails["ids"].append(event.id)
+                thumbnails["images"].append(img)
+                thumbnails["metadatas"].append(metadata)
+                if event.data.get("description") is not None:
+                    descriptions["ids"].append(event.id)
+                    descriptions["documents"].append(event.data["description"])
+                    descriptions["metadatas"].append(metadata)
+
+            self.thumbnail.upsert(
+                images=thumbnails["images"],
+                metadatas=thumbnails["metadatas"],
+                ids=thumbnails["ids"],
+            )
+            self.description.upsert(
+                documents=descriptions["documents"],
+                metadatas=descriptions["metadatas"],
+                ids=descriptions["ids"],
+            )
+
+            logger.info(
+                "Embedded %d thumbnails and %d descriptions in %s seconds",
+                len(thumbnails["ids"]),
+                len(descriptions["ids"]),
+                time.time() - st,
+            )
+
+            os.remove(f"{CONFIG_DIR}/.reindex_events")

frigate/embeddings/processor.py

@@ -10,18 +10,13 @@ from multiprocessing.synchronize import Event as MpEvent
 
 import google.generativeai as genai
 import numpy as np
-from chromadb import Collection
-from chromadb import HttpClient as ChromaClient
-from chromadb.config import Settings
 from peewee import DoesNotExist
 from PIL import Image
-from playhouse.shortcuts import model_to_dict
 
 from frigate.config import FrigateConfig
 from frigate.models import Event
 
-from .functions.clip import ClipEmbedding
-from .functions.minilm_l6_v2 import MiniLMEmbedding
+from . import Embeddings, get_metadata
 
 logger = logging.getLogger(__name__)
 
@@ -32,39 +27,25 @@ class EmbeddingProcessor(threading.Thread):
     def __init__(
         self,
         config: FrigateConfig,
+        embeddings: Embeddings,
         queue: Queue,
         stop_event: MpEvent,
     ) -> None:
         threading.Thread.__init__(self)
         self.name = "chroma"
         self.config = config
+        self.embeddings = embeddings
         self.queue = queue
         self.stop_event = stop_event
-        self.chroma: ChromaClient = None
-        self.thumbnail: Collection = None
-        self.description: Collection = None
         self.gemini: genai.GenerativeModel = None
 
     def run(self) -> None:
         """Maintain a Chroma vector database for semantic search."""
 
         # Exit if disabled
-        if not self.config.semantic_search.enabled:
+        if self.embeddings is None:
             return
 
-        # Create the database
-        self.chroma = ChromaClient(
-            host="127.0.0.1", settings=Settings(anonymized_telemetry=False)
-        )
-
-        # Create/Load the collection(s)
-        self.thumbnail = self.chroma.get_or_create_collection(
-            name="event_thumbnail", embedding_function=ClipEmbedding()
-        )
-        self.description = self.chroma.get_or_create_collection(
-            name="event_description", embedding_function=MiniLMEmbedding()
-        )
-
         ## Initialize Gemini
         if self.config.gemini.enabled:
             genai.configure(api_key=self.config.gemini.api_key)
@@ -88,13 +69,7 @@ class EmbeddingProcessor(threading.Thread):
                 continue
 
             # Extract valid event metadata
-            metadata = {
-                k: v
-                for k, v in model_to_dict(event).items()
-                if k not in ["id", "thumbnail"]
-                and v is not None
-                and isinstance(v, (str, int, float, bool))
-            }
+            metadata = get_metadata(event)
             thumbnail = base64.b64decode(event.thumbnail)
 
             # Encode the thumbnail
@@ -124,7 +99,7 @@ class EmbeddingProcessor(threading.Thread):
 
         # Encode the thumbnail
         img = np.array(Image.open(io.BytesIO(thumbnail)).convert("RGB"))
-        self.thumbnail.add(
+        self.embeddings.thumbnail.upsert(
             images=[img],
             metadatas=[metadata],
             ids=[event_id],
@@ -163,7 +138,7 @@ class EmbeddingProcessor(threading.Thread):
         event.save()
 
         # Encode the description
-        self.description.add(
+        self.embeddings.description.upsert(
             documents=[description],
             metadatas=[metadata],
             ids=[event.id],

frigate/events/cleanup.py

@@ -10,6 +10,7 @@ from pathlib import Path
 
 from frigate.config import FrigateConfig
 from frigate.const import CLIPS_DIR
+from frigate.embeddings import Embeddings
 from frigate.models import Event, Timeline
 
 logger = logging.getLogger(__name__)
@@ -21,10 +22,13 @@ class EventCleanupType(str, Enum):
 
 
 class EventCleanup(threading.Thread):
-    def __init__(self, config: FrigateConfig, stop_event: MpEvent):
+    def __init__(
+        self, config: FrigateConfig, embeddings: Embeddings, stop_event: MpEvent
+    ):
         threading.Thread.__init__(self)
         self.name = "event_cleanup"
         self.config = config
+        self.embeddings = embeddings
         self.stop_event = stop_event
         self.camera_keys = list(self.config.cameras.keys())
         self.removed_camera_labels: list[str] = None
@@ -204,11 +208,12 @@ class EventCleanup(threading.Thread):
             media_path = Path(f"{os.path.join(CLIPS_DIR, media_name)}-clean.png")
             media_path.unlink(missing_ok=True)
 
-        (
-            Event.delete()
-            .where(Event.id << [event.id for event in duplicate_events])
-            .execute()
-        )
+        duplicate_ids = [event.id for event in duplicate_events]
+        Event.delete().where(Event.id << duplicate_ids).execute()
+        # Also remove from embeddings database
+        if self.embeddings is not None and len(duplicate_ids) > 0:
+            self.embeddings.thumbnail.delete(duplicate_ids)
+            self.embeddings.description.delete(duplicate_ids)
 
     def run(self) -> None:
         # only expire events every 5 minutes
@@ -223,10 +228,20 @@ class EventCleanup(threading.Thread):
             self.expire(EventCleanupType.snapshots)
             self.purge_duplicates()
 
-            # drop events from db where has_clip and has_snapshot are false
-            delete_query = Event.delete().where(
-                Event.has_clip == False, Event.has_snapshot == False
+            # get list of ids that have both expired clips and snapshots
+            # so we can delete them from the embeddings db (and the events table)
+            events = (
+                Event.select(Event.id)
+                .where(Event.has_clip == False, Event.has_snapshot == False)
+                .iterator()
             )
-            delete_query.execute()
+            events_to_delete = [e.id for e in events]
+
+            if self.embeddings is not None and len(events_to_delete) > 0:
+                self.embeddings.thumbnail.delete(events_to_delete)
+                self.embeddings.description.delete(events_to_delete)
+
+            # drop events from db where has_clip and has_snapshot are false
+            Event.delete().where(Event.id << events_to_delete).execute()
 
         logger.info("Exiting event cleanup...")

frigate/http.py

@@ -1,6 +1,7 @@
 import base64
 import copy
 import glob
+import io
 import json
 import logging
 import os
@@ -18,8 +19,6 @@ import numpy as np
 import pytz
 import requests
 from chromadb import Collection, QueryResult
-from chromadb import HttpClient as ChromaClient
-from chromadb.config import Settings
 from flask import (
     Blueprint,
     Flask,
@@ -31,6 +30,7 @@ from flask import (
     request,
 )
 from peewee import DoesNotExist, fn, operator
+from PIL import Image
 from playhouse.shortcuts import model_to_dict
 from playhouse.sqliteq import SqliteQueueDatabase
 from tzlocal import get_localzone_name
@@ -45,8 +45,7 @@ from frigate.const import (
     MAX_SEGMENT_DURATION,
     RECORD_DIR,
 )
-from frigate.embeddings.functions.clip import ClipEmbedding
-from frigate.embeddings.functions.minilm_l6_v2 import MiniLMEmbedding
+from frigate.embeddings import Embeddings
 from frigate.events.external import ExternalEventProcessor
 from frigate.models import Event, Previews, Recordings, Regions, Timeline
 from frigate.object_processing import TrackedObject
@@ -79,6 +78,7 @@ def create_app(
     onvif: OnvifController,
     external_processor: ExternalEventProcessor,
     plus_api: PlusApi,
+    embeddings: Embeddings,
 ):
     app = Flask(__name__)
 
@@ -108,15 +108,7 @@ def create_app(
     app.plus_api = plus_api
     app.camera_error_image = None
     app.hwaccel_errors = []
-    app.chroma = ChromaClient(
-        host="127.0.0.1", settings=Settings(anonymized_telemetry=False)
-    )
-    app.thumbnail_collection = app.chroma.get_or_create_collection(
-        name="event_thumbnail", embedding_function=ClipEmbedding()
-    )
-    app.description_collection = app.chroma.get_or_create_collection(
-        name="event_description", embedding_function=MiniLMEmbedding()
-    )
+    app.embeddings = embeddings
 
     app.register_blueprint(bp)
 
@@ -529,6 +521,9 @@ def delete_event(id):
 
     event.delete_instance()
     Timeline.delete().where(Timeline.source_id == id).execute()
+    if current_app.embeddings is not None:
+        current_app.embeddings.thumbnail.delete(ids=[id])
+        current_app.embeddings.description.delete(ids=[id])
     return make_response(
         jsonify({"success": True, "message": "Event " + id + " deleted"}), 200
     )
@@ -1013,6 +1008,7 @@ def events():
     min_length = request.args.get("min_length", type=float)
     max_length = request.args.get("max_length", type=float)
     search = request.args.get("search", type=str) or None
+    like = request.args.get("like", type=str) or None
 
     clauses = []
 
@@ -1168,37 +1164,57 @@ def events():
 
     # Handle semantic search
     event_order = None
-    if search is not None:
+    if current_app.embeddings is not None:
         where = None
         if len(embeddings_filters) > 1:
             where = {"$and": embeddings_filters}
         elif len(embeddings_filters) == 1:
             where = embeddings_filters[0]
 
-        # Grab the ids of the events that match based on CLIP embeddings
-        thumbnails: Collection = current_app.thumbnail_collection
-        thumb_result: QueryResult = thumbnails.query(
-            query_texts=[search],
-            n_results=int(limit),
-            where=where,
-        )
-        thumb_ids = dict(zip(thumb_result["ids"][0], thumb_result["distances"][0]))
+        if like is not None:
+            # Grab the ids of events that match the thumbnail image embeddings
+            thumbnails: Collection = current_app.embeddings.thumbnail
+            search_event = Event.get(Event.id == like)
+            thumbnail = base64.b64decode(search_event.thumbnail)
+            img = np.array(Image.open(io.BytesIO(thumbnail)).convert("RGB"))
+            thumb_result: QueryResult = thumbnails.query(
+                query_images=[img],
+                n_results=int(limit),
+                where=where,
+            )
+            event_order = dict(
+                zip(thumb_result["ids"][0], thumb_result["distances"][0])
+            )
 
-        # Grab the ids of the events that match based on MiniLM embeddings
-        descriptions: Collection = current_app.description_collection
-        desc_result: QueryResult = descriptions.query(
-            query_texts=[search],
-            n_results=int(limit),
-            where=where,
-        )
-        desc_ids = dict(zip(desc_result["ids"][0], desc_result["distances"][0]))
+            # For like, we want to remove all other filters
+            clauses = [(Event.id << list(event_order.keys()))]
 
-        event_order = {
-            k: min(i for i in (thumb_ids.get(k), desc_ids.get(k)) if i is not None)
-            for k in thumb_ids.keys() | desc_ids
-        }
+        elif search is not None:
+            # Grab the ids of the events that match based on CLIP embeddings
+            thumbnails: Collection = current_app.embeddings.thumbnail
+            thumb_result: QueryResult = thumbnails.query(
+                query_texts=[search],
+                n_results=int(limit),
+                where=where,
+            )
+            thumb_ids = dict(zip(thumb_result["ids"][0], thumb_result["distances"][0]))
 
-        clauses.append((Event.id << list(event_order.keys())))
+            # Grab the ids of the events that match based on MiniLM embeddings
+            descriptions: Collection = current_app.embeddings.description
+            desc_result: QueryResult = descriptions.query(
+                query_texts=[search],
+                n_results=int(limit),
+                where=where,
+            )
+            desc_ids = dict(zip(desc_result["ids"][0], desc_result["distances"][0]))
+
+            event_order = {
+                k: min(i for i in (thumb_ids.get(k), desc_ids.get(k)) if i is not None)
+                for k in thumb_ids.keys() | desc_ids
+            }
+
+            # For search, we want to keep all the other clauses and filters
+            clauses.append((Event.id << list(event_order.keys())))
 
     events = (
         Event.select(*selected_columns)

web-old/src/routes/Events.jsx

@@ -259,6 +259,16 @@ export default function Events({ path, ...props }) {
     setState({ ...state, showDownloadMenu: true });
   };
 
+  const showSimilarEvents = (event_id, e) => {
+    if (e) {
+      e.stopPropagation();
+    }
+    if (searchParams?.like == event_id) {
+      return;
+    }
+    onFilter('like', event_id);
+  };
+
   const showSubmitToPlus = (event_id, label, box, e) => {
     if (e) {
       e.stopPropagation();
@@ -289,6 +299,7 @@ export default function Events({ path, ...props }) {
     (name, value) => {
       setShowInProgress(false);
       const updatedParams = { ...searchParams, [name]: value };
+      if (name !== 'like') delete updatedParams['like'];
       setSearchParams(updatedParams);
       const queryString = Object.keys(updatedParams)
         .map((key) => {
@@ -311,7 +322,10 @@ export default function Events({ path, ...props }) {
     onFilter('is_submitted', searchParams.is_submitted);
   }, [searchParams, onFilter]);
 
-  const isDone = (eventPages?.[eventPages.length - 1]?.length ?? 0) < API_LIMIT;
+  const isDone =
+    (eventPages?.[eventPages.length - 1]?.length ?? 0) < API_LIMIT ||
+    (searchParams?.search?.length ?? 0) > 0 ||
+    (searchParams?.like?.length ?? 0) > 0;
 
   // hooks for infinite scroll
   const observer = useRef();
@@ -471,7 +485,8 @@ export default function Events({ path, ...props }) {
               download
             />
           )}
-          {(event?.data?.type || 'object') == 'object' &&
+          {config.plus.enabled &&
+            (event?.data?.type || 'object') == 'object' &&
             downloadEvent.end_time &&
             downloadEvent.has_snapshot &&
             !downloadEvent.plus_id && (
@@ -482,7 +497,7 @@ export default function Events({ path, ...props }) {
                 onSelect={() => showSubmitToPlus(downloadEvent.id, downloadEvent.label, downloadEvent.box)}
               />
             )}
-          {downloadEvent.plus_id && (
+          {config.plus.enabled && downloadEvent.plus_id && (
             <MenuItem
               icon={UploadPlus}
               label={'Sent to Frigate+'}
@@ -710,6 +725,7 @@ export default function Events({ path, ...props }) {
                       });
                     }}
                     onSave={onSave}
+                    showSimilarEvents={showSimilarEvents}
                     showSubmitToPlus={showSubmitToPlus}
                   />
                 );
@@ -750,6 +766,7 @@ export default function Events({ path, ...props }) {
                     });
                   }}
                   onSave={onSave}
+                  showSimilarEvents={showSimilarEvents}
                   showSubmitToPlus={showSubmitToPlus}
                 />
               );
@@ -782,6 +799,7 @@ function Event({
   onDownloadClick,
   onReady,
   onSave,
+  showSimilarEvents,
   showSubmitToPlus,
 }) {
   const apiHost = useApiHost();
@@ -851,30 +869,39 @@ function Event({
             </div>
           </div>
           <div class="hidden sm:flex flex-col justify-end mr-2">
-            {event.end_time && event.has_snapshot && (event?.data?.type || 'object') == 'object' && (
-              <Fragment>
-                {event.plus_id ? (
-                  <div className="uppercase text-xs underline">
-                    <Link
-                      href={`https://plus.frigate.video/dashboard/edit-image/?id=${event.plus_id}`}
-                      target="_blank"
-                      rel="nofollow"
-                    >
-                      Edit in Frigate+
-                    </Link>
-                  </div>
-                ) : (
-                  <Button
-                    color="gray"
-                    disabled={uploading.includes(event.id)}
-                    onClick={(e) => showSubmitToPlus(event.id, event.label, event?.data?.box || event.box, e)}
-                  >
-                    {uploading.includes(event.id) ? 'Uploading...' : 'Send to Frigate+'}
-                  </Button>
-                )}
-              </Fragment>
+            {event.id && (
+              <Button color="gray" onClick={(e) => showSimilarEvents(event.id, e)}>
+                Find Similar
+              </Button>
             )}
           </div>
+          {config.plus.enabled && (
+            <div class="hidden sm:flex flex-col justify-end mr-2">
+              {event.end_time && event.has_snapshot && (event?.data?.type || 'object') == 'object' && (
+                <Fragment>
+                  {event.plus_id ? (
+                    <div className="uppercase text-xs underline">
+                      <Link
+                        href={`https://plus.frigate.video/dashboard/edit-image/?id=${event.plus_id}`}
+                        target="_blank"
+                        rel="nofollow"
+                      >
+                        Edit in Frigate+
+                      </Link>
+                    </div>
+                  ) : (
+                    <Button
+                      color="gray"
+                      disabled={uploading.includes(event.id)}
+                      onClick={(e) => showSubmitToPlus(event.id, event.label, event?.data?.box || event.box, e)}
+                    >
+                      {uploading.includes(event.id) ? 'Uploading...' : 'Send to Frigate+'}
+                    </Button>
+                  )}
+                </Fragment>
+              )}
+            </div>
+          )}
           <div class="flex flex-col">
             <Delete
               className="h-6 w-6 cursor-pointer"